Very Short-Term Solar Irradiance Forecasting using Multilayered Long-Short Term Memory

The intermittent nature of the resources (inputs: sunlight) results in intermittent energy output (power) from the photovoltaic energy systems. Therefore, accurate prediction of the photovoltaic output is a key area of research across the globe. Direct normal irradiance
$DNI$
) forecasts are essential for the planning of a distributed energy
$DE$
) powered facility or concentrated solar power systems. Some areas are more prone to have random weather conditions, one such city is Melbourne, Australia. This makes the forecasting a challenging task for conventional approaches and ther is a need for more sophisticated algorithms to makes accurate predictions. Accurate forecasting aids in optimally scheduling of HVAC loads, demand side management, trading of transactive energy, asset optimization and data analytics. In this manuscript a faster, efficient and less exhaustive algorithm for quarter-hour (15 minutes) forecasts for sites prone to weather irregularities is presented. This study discusses a case study using a machine learning based multi-layered long-short term memory neural networks
$(LSTM-NN$
) with Adam optimizer are used to forecast
$DNI$
. This research addresses the best forecasting for subtropical conditions specifically and evaluates its performance by comparing with the most used methods including
$ARIMA,\ ARMA,\ SVR$
, Persistence, and
$AR$
. The proposed nested
$LSTM$
model was assessed by comparing
$RMSE(27.35),\ MAE(6.879)$
and
$MBE(0.42)$
values with other aforementioned methods. The results reveal the optimal performance of
$LSTM$
with reduced operational time, improved precision, and accuracy.